different shape size using Invertd or transforms.inverse

[sherjy]

Hiii, I used Invertd and transforms.inverse in monai 0.6.0 for inversing preprocessed transformations. However, it seems it cannot inverse the transforms correctly.
The codes and shape results are as follows:

temp = {
        'now': os.path.join(temp_path, 'data.nii.gz'),
        'now_seg': os.path.join(temp_path, 'seg.nii.gz'),
        }
data_dict.append(temp)

# def preprocessed transforms
transforms = Compose([])
transforms.transforms += (LoadImaged(keys=['now', 'now_seg']),)
transforms.transforms += (AddChanneld(keys=['now', 'now_seg']),)
transforms.transforms += (CropForegroundd(keys=['now', 'now_seg'], source_key='now',
                                          select_fn=lambda x: x > pp['vmin'], margin=pp['vmax']),) # attention here
transforms.transforms += (LongBoundaryPadd(keys=['now', 'now_seg']),)
transforms.transforms += (SpacingResized(['now', 'now_seg'],
                                         [pp['height'], pp['height'], -1]),)
if 'cpresample' in pp:
    transforms.transforms += (CopyResampled(keys=['now', 'now_seg'],
                                            slice_thick=pp['cpresample']),)
transforms.transforms += (ExtractSliced(keys=['now', 'now_seg'],
                                        slice_thick=pp['st_range'][0]),)
transforms.transforms += (ScaleIntensityRanged(keys=['now'], a_min=pp['win_lower'],
                                               a_max=pp['win_upper'],
                                               b_min=pp['vmin'],
                                               b_max=pp['vmax'], clip=True),)
transforms.transforms += (ResizeWithPadOrCropd(['now', 'now_seg'],
                                               spatial_size=[pp['height'], pp['height'], pp['slicenum']]),)
transforms.transforms += (ToTensord(keys=['now', 'now_seg']),)

# define post transforms
post_transforms = Compose([
    Activationsd(keys="now", sigmoid=True),
    Invertd(
        keys="now",  # invert the `pred` data field, also support multiple fields
        transform=transforms,
        orig_keys="now",  # get the previously applied pre_transforms information on the `img` data field,
    ),
])

# print original image shape
load_data = LoadImaged(keys=['now', 'now_seg'])
origin_data = load_data(data_dict[0])
print(origin_data['now'].shape) #(512, 512, 60)
print('end_0')

# print transformed image shape
transformed_data = transforms(data_dict[0])
print(transformed_data['now'].shape) # torch.Size([1, 128, 128, 64]) 
print('end_1')

# print inverse transformed image shape
img_dict = {"now": transformed_data['now'], "now_transforms": transformed_data["now_transforms"]}
output = post_transforms(img_dict)
print(output['now'].shape) #torch.Size([1, 194, 257, 66]) 
print('end_2')

Original input image size is (512,512,60). After preprocessed transformations, the shape is (1,128,128,64). However, the inversed results size is (1,194,257,66). I used both Invertd to create a transforms composition or transforms.inverse, both the inversed methods cannot get correct original shape.

[Nic-Ma]

Hi @sherjy ,

Thanks for your interest and experiments with MONAI.
I see you have several customized transforms, like: SpacingResized, have you implemented the inverse operation for them properly? If it doesn't have inverse logic, then the test result is expected I think.

Thanks.

[sherjy]

Thanks a lot for reminding me this! Will try!!